Enhancing Thyroid Ultrasound Image Classification with Deep Transfer Learning

Ultrasound imaging is the primary diagnostic tool for thyroid nodules. However, clinical decisions regarding these nodules often involve subjective judgments influenced by individual experience and practice variability. To improve the objectivity and accuracy of these decisions, this study proposes a VGG-16-based deep learning model for accurately predicting the benign or malignant nature of thyroid nodules. The objective was to classify thyroid ultrasound images into two categories: benign and malignant. To achieve this, a dataset of thyroid ultrasound images was collected from The Cancer Genome Atlas Thyroid Cancer (TCGA-THCA) repository and the Digital Database of Thyroid Ultrasound Images (DDTI). To augment the dataset and mitigate overfitting, histogram equalization was employed. The proposed model leverages the pre-trained VGG-16 architecture, which was fine-tuned to achieve optimal performance in distinguishing benign and malignant thyroid nodules. This study focuses on optimizing model parameters and architecture to enhance the efficiency of the proposed approach. For the assessment of the performance of the proposed DL model, accuracy, sensitivity, recall, intersection over union (IoU), specificity, and F1-score will be calculated in the study. The results will also be compared with other classification methods that used transfer learning to determine the degree of deep learning in classifying liver tumors. Our research yielded the greatest possible results, which we got, reaching 99.80% accuracy, 100% precision, F1- scare 98.9%, and a recall of 98%.